The target sits on the right. At every node the model proposes 2–3 candidate disconnections — retrosynthetic steps that break one bond to reveal a simpler precursor — each tagged with a predicted success probability, the way models trained on millions of literature reactions (AiZynthFinder, IBM RXN) score real disconnections.
A Monte Carlo tree search — the same idea AlphaGo used to search game trees — repeatedly walks down from the target using an upper-confidence rule, reaches a candidate route, and backs the result up the path:
select = argmax( avgValue + C·√(ln(N_parent)/N_child) )
Routes that end in cheap, purchasable building blocks with high cumulative disconnection probability get reinforced; weak branches get visited less. After the search budget is spent, the most-visited path from target to building blocks is highlighted gold.